Methodology
Your approved brand information is the reference. Observed AI answers are the evidence. Three separate metrics help you understand the difference.
BrandVault holds the facts, products, audience and positioning your brand approves. Brand Index helps agents discover and retrieve brand information, distinguishing sourced records from owner-approved records.
Measurement asks a separate question: how do tested AI answers describe and recommend the brand? An approved record is the reference for that comparison, not independent proof that every claim is true or that a model will retrieve it.
Branded questions name the business. Unbranded questions ask about a category or need without naming it. The report keeps those contexts distinct.
The AI Visibility report
Does the answer mention your brand?
For each eligible unbranded question, calculate the share of successfully judged answers that mention the brand. Then average those question-level shares. Brand mentions can include the name, an alias or the domain.
The app shows the total mention count and total eligible answer count, such as 2 of 28 unbranded answers mention you.
Does the description align with your positioning?
The answer judge assesses alignment on a scale from 0 to 1 when positioning criteria are supplied. Average the available alignment scores within each question, then average those question-level scores. Both branded and unbranded answers can contribute when scored.
The app displays the result as a percentage, such as Your alignment score was 45%. This is an alignment score, not a claim that 45% of all facts are correct.
Is your brand recommended first?
The current calculation counts eligible unbranded answers where the judge marks the brand as recommended first. All successfully judged answers to those questions remain in the denominator, including answers that do not recommend or rank brands. Calculate a share for each question, then average the shares.
The app labels this count as answers that “recommended you,” such as 0 of 28 unbranded answers recommended you. A mention or a lower-ranked recommendation alone does not count toward this metric.
Visibility and Endorsement Strength exclude entity-resolution questions: questions used to distinguish similarly named businesses can introduce the brand name themselves.
Each eligible question has equal weight. If some questions have fewer successful answers, the average question-level score can differ from dividing the report’s total numerator by its total denominator.
These fixed bands describe the metric value. They are not confidence levels, statistical significance or percentile rankings against other brands.
Provider failures and answers without a successful judgment are excluded from metric calculations and listed as exclusions. Positioning Strength also excludes answers without an alignment score. Missing results are not scored as zero.
If a metric has no eligible results, it is not measured that run. A partial report describes only the available results; it cannot establish how a missing model sees the brand.
Summary brings together the narrative, three metrics and exclusions. Question Analysis lets you inspect the tested questions and recorded answers. Sources shows the citations available from the answers.
Use the report version, dates, question set, model coverage and answer counts to understand the sample. Scores reflect automated judgments of observed answers, which can be imperfect. Review the evidence before acting.
Recommendations propose next steps with an explanation and, where available, suggested copy. You review the action and decide what to change. A recommendation is not a completed change or a promised outcome.
Explore the report and recommendationsAfter updating your brand information or other source material, a later report can show whether the tested answers changed. Compare the questions, model coverage, scoring method and approved brand information before interpreting a difference.
When the BrandVault changes, the reference used to assess alignment can change too. A higher score alone does not establish that the same answers improved.
Model updates, other website changes and competitor activity can also affect results. AI answers do not establish what all buyers believe or prove a change caused traffic, sales or revenue.
Likewise, an agent requesting a Brand Index record shows retrieval activity. It does not by itself prove citation, recommendation or a customer visit.
Claim a free BrandVault, approve your information and explore what the reports show.
Retrieval observation is separate from the controlled tests above. Our observation window is activity reaching Optimly’s index, not private conversations or the full market.
Current crawl-health analytics group requests by response status and user-agent labels. User-agent labels indicate a claimed client identity, not verified buyer intent. Worker origin fetches are excluded to avoid counting that internal hop as another client request; this is not a count of unique buyers or a general retry-deduplication method.
A page fetch alone does not reveal a buyer question. In a scoped review, we distinguish explicit ask-like information from inferred intent, including identifiable asks that remain unresolved or return 404/410. An HTTP 200 alone does not establish that a need was answered.
Any request analysis should state its source, date window, attribution confidence and counting method. We compare relevant BrandVault facts and their date with the available evidence, identify unknowns and consider plausible fit or mismatch. There is no standardized demand or buyer-fit score in the model-test report.
Updating BrandVault can change the context for an interpretation. It does not change historical requests or prove new demand. Explore the brand workflow.
